Explainable AI for Medical Diagnoses: Advancing Explainability and Assurance in AI-Supported Healthcare

Patil, Hitesh Vijay (2025) Explainable AI for Medical Diagnoses: Advancing Explainability and Assurance in AI-Supported Healthcare. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (378kB) | Preview

Abstract

In recent years, the integration of Artificial Intelligence (AI) into the healthcare sector has experienced significant growth, particularly in the field of diagnostic imaging. These innovations hold promise for enhancing diagnostic accuracy and efficiency; however, concerns regarding fairness and model transparency have become increasingly important. This study addresses two fundamental aspects of trustworthy AI in healthcare: bias mitigation and post hoc interpretability, specifically in the context of pneumonia classification using chest radiography images. The project employs fairness-aware learning techniques, including Reweighing and Adversarial Debiasing from the AI Fairness 360 (AIF360) toolkit, to mitigate gender bias in prediction outcomes. Concurrently, interpretability methods such as Grad-CAM, SHAP, and LIME were utilized to generate both visual and feature-level explanations of the model behavior. A relatively shallow Convolutional Neural Network (CNN) was used as the baseline model, selected because of dataset size constraints and limited computational resources. Although the model achieved high accuracy on the training set, including instances of perfect classification, this raised concerns about potential overfitting. Fairness metrics prior to the intervention indicated considerable bias, with Disparate Impact and Statistical Parity Difference deviating from the ideal thresholds. Post-mitigation evaluations demonstrated improved fairness, albeit with minor performance trade-offs. The explainability outputs provided insights into the model’s decision-making logic and helped identify the underlying biases associated with sensitive attributes. Given the limited dataset and simple model architecture, this study was framed as a proof of concept rather than a robust deployment-ready solution. These findings underscore the feasibility of integrating fairness and interpretability techniques into medical AI pipelines, while highlighting the need for further research using deeper architectures, such as ResNet or DenseNet, and larger, more representative datasets to support generalizable outcomes.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Raj, Kislay UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence R Medicine > Healthcare Industry
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	12 Aug 2026 09:25
Last Modified:	12 Aug 2026 09:25
URI:	https://norma.ncirl.ie/id/eprint/9511

Actions (login required)

View Item